diff --git a/codes/models/archs/RRDBNet_arch.py b/codes/models/archs/RRDBNet_arch.py index 558c24c9..7be467d5 100644 --- a/codes/models/archs/RRDBNet_arch.py +++ b/codes/models/archs/RRDBNet_arch.py @@ -3,19 +3,20 @@ import torch import torch.nn as nn import torch.nn.functional as F import models.archs.arch_util as arch_util +from models.archs.arch_util import PixelUnshuffle import torchvision import switched_conv as switched_conv class ResidualDenseBlock_5C(nn.Module): - def __init__(self, nf=64, gc=32, bias=True): + def __init__(self, nf=64, gc=32, bias=True, late_stage_kernel_size=3, late_stage_padding=1): super(ResidualDenseBlock_5C, self).__init__() # gc: growth channel, i.e. intermediate channels self.conv1 = nn.Conv2d(nf, gc, 3, 1, 1, bias=bias) self.conv2 = nn.Conv2d(nf + gc, gc, 3, 1, 1, bias=bias) - self.conv3 = nn.Conv2d(nf + 2 * gc, gc, 3, 1, 1, bias=bias) - self.conv4 = nn.Conv2d(nf + 3 * gc, gc, 3, 1, 1, bias=bias) - self.conv5 = nn.Conv2d(nf + 4 * gc, nf, 3, 1, 1, bias=bias) + self.conv3 = nn.Conv2d(nf + 2 * gc, gc, late_stage_kernel_size, 1, late_stage_padding, bias=bias) + self.conv4 = nn.Conv2d(nf + 3 * gc, gc, late_stage_kernel_size, 1, late_stage_padding, bias=bias) + self.conv5 = nn.Conv2d(nf + 4 * gc, nf, late_stage_kernel_size, 1, late_stage_padding, bias=bias) self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True) # initialization @@ -32,9 +33,15 @@ class ResidualDenseBlock_5C(nn.Module): # Multiple 5-channel residual block that uses learned switching to diversify its outputs. +# If multi_head_input=False: takes standard (b,f,w,h) input tensor; else takes (b,heads,f,w,h) input tensor. Note that the default RDB block does not support this format, so use SwitchedRDB_5C_MultiHead for this case. +# If collapse_heads=True, outputs (b,f,w,h) tensor. +# If collapse_heads=False, outputs (b,heads,f,w,h) tensor. class SwitchedRDB_5C(switched_conv.MultiHeadSwitchedAbstractBlock): - def __init__(self, nf=64, gc=32, num_convs=8, num_heads=2, init_temperature=1): - rdb5c = functools.partial(ResidualDenseBlock_5C, nf, gc) + def __init__(self, nf=64, gc=32, num_convs=8, num_heads=2, init_temperature=1, multi_head_input=False, collapse_heads=True, force_block=None): + if force_block is None: + rdb5c = functools.partial(ResidualDenseBlock_5C, nf, gc) + else: + rdb5c = force_block super(SwitchedRDB_5C, self).__init__( rdb5c, nf, @@ -43,22 +50,70 @@ class SwitchedRDB_5C(switched_conv.MultiHeadSwitchedAbstractBlock): att_kernel_size=3, att_pads=1, initial_temperature=init_temperature, + multi_head_input=multi_head_input, + concat_heads_into_filters=collapse_heads, ) - self.mhead_collapse = nn.Conv2d(num_heads * nf, nf, 1) + self.collapse_heads = collapse_heads + if self.collapse_heads: + self.mhead_collapse = nn.Conv2d(num_heads * nf, nf, 1) + arch_util.initialize_weights([self.mhead_collapse], 1) arch_util.initialize_weights([sw.attention_conv1 for sw in self.switches] + - [sw.attention_conv2 for sw in self.switches] + - [self.mhead_collapse], 1) + [sw.attention_conv2 for sw in self.switches], 1) def forward(self, x, output_attention_weights=False): outs = super(SwitchedRDB_5C, self).forward(x, output_attention_weights) if output_attention_weights: outs, atts = outs - # outs need to be collapsed back down to a single heads worth of data. - out = self.mhead_collapse(outs) + + if self.collapse_heads: + # outs need to be collapsed back down to a single heads worth of data. + out = self.mhead_collapse(outs) + else: + out = outs + return out, atts +# Implementation of ResidualDenseBlock_5C which compresses multiple switching heads via a Conv3d before doing RDB +# computation. +class ResidualDenseBlock_5C_WithMheadConverter(ResidualDenseBlock_5C): + def __init__(self, nf=64, gc=32, bias=True, heads=2): + # Switched blocks generally operate at low resolution, kernel size is much less important, therefore set to 1. + super(ResidualDenseBlock_5C_WithMheadConverter, self).__init__(nf=nf, gc=gc, bias=bias, late_stage_kernel_size=1, + late_stage_padding=0) + self.heads = heads + self.converter = nn.Conv3d(nf, nf, kernel_size=(heads, 1, 1), stride=(heads, 1, 1)) + + # Accepts input of shape (b, heads, f, w, h) + def forward(self, x): + # Permute filter dim to 1. + x = x.permute(0, 2, 1, 3, 4) + x = self.converter(x) + x = torch.squeeze(x, dim=2) + return super(ResidualDenseBlock_5C_WithMheadConverter, self).forward(x) + + +# Multiple 5-channel residual block that uses learned switching to diversify its outputs. The difference between this +# block and SwitchedRDB_5C is this block accepts multi-headed inputs of format (b,heads,f,w,h). +# +# It does this by performing a Conv3d on the first block, which convolves all heads and collapses them to a dimension +# of 1. The tensor is then squeezed and performs identically to SwitchedRDB_5C from there. +class SwitchedRDB_5C_MultiHead(SwitchedRDB_5C): + def __init__(self, nf=64, gc=32, num_convs=8, num_heads=2, init_temperature=1, collapse_heads=False): + rdb5c = functools.partial(ResidualDenseBlock_5C_WithMheadConverter, nf, gc, heads=num_heads) + super(SwitchedRDB_5C_MultiHead, self).__init__( + nf=nf, + gc=gc, + num_convs=num_convs, + num_heads=num_heads, + init_temperature=init_temperature, + multi_head_input=True, + collapse_heads=collapse_heads, + force_block=rdb5c, + ) + + class RRDB(nn.Module): '''Residual in Residual Dense Block''' @@ -74,13 +129,26 @@ class RRDB(nn.Module): out = self.RDB3(out) return out * 0.2 + x + +class LowDimRRDB(RRDB): + def __init__(self, nf, gc=32, dimensional_adjustment=4): + super(LowDimRRDB, self).__init__(nf * (dimensional_adjustment ** 2), gc * (dimensional_adjustment ** 2)) + self.unshuffle = PixelUnshuffle(dimensional_adjustment) + self.shuffle = nn.PixelShuffle(dimensional_adjustment) + + def forward(self, x): + x = self.unshuffle(x) + x = super(LowDimRRDB, self).forward(x) + return self.shuffle(x) + + # RRDB block that uses switching on the individual RDB modules that compose it to increase learning diversity. class SwitchedRRDB(RRDB): - def __init__(self, nf, gc=32, num_convs=8, init_temperature=1, final_temperature_step=1): - super(RRDB, self).__init__() - self.RDB1 = SwitchedRDB_5C(nf, gc, num_convs=num_convs, init_temperature=init_temperature) - self.RDB2 = SwitchedRDB_5C(nf, gc, num_convs=num_convs, init_temperature=init_temperature) - self.RDB3 = SwitchedRDB_5C(nf, gc, num_convs=num_convs, init_temperature=init_temperature) + def __init__(self, nf, gc=32, num_convs=8, init_temperature=1, final_temperature_step=1, switching_block=SwitchedRDB_5C): + super(SwitchedRRDB, self).__init__(nf, gc) + self.RDB1 = switching_block(nf, gc, num_convs=num_convs, init_temperature=init_temperature) + self.RDB2 = switching_block(nf, gc, num_convs=num_convs, init_temperature=init_temperature) + self.RDB3 = switching_block(nf, gc, num_convs=num_convs, init_temperature=init_temperature) self.init_temperature = init_temperature self.final_temperature_step = final_temperature_step self.running_mean = 0 @@ -116,6 +184,53 @@ class SwitchedRRDB(RRDB): self.running_mean = 0 return val + +# Identical to LowDimRRDB but wraps an RRDB rather than inheriting from it. TODO: remove LowDimRRDB when backwards +# compatibility is no longer desired. +class LowDimRRDBWrapper(nn.Module): + # Do not specify nf or gc on the partial_rrdb passed in. That will be done by the wrapper. + def __init__(self, nf, partial_rrdb, gc=32, dimensional_adjustment=4): + super(LowDimRRDBWrapper, self).__init__() + self.rrdb = partial_rrdb(nf=nf * (dimensional_adjustment ** 2), gc=gc * (dimensional_adjustment ** 2)) + self.unshuffle = PixelUnshuffle(dimensional_adjustment) + self.shuffle = nn.PixelShuffle(dimensional_adjustment) + + def forward(self, x): + x = self.unshuffle(x) + x = self.rrdb(x) + return self.shuffle(x) + +# RRDB block that uses multi-headed switching on multiple individual RDB blocks to improve diversity. Multiple heads +# are annealed internally. This variant has a depth of 4 RDB blocks, rather than 3 like others above. +class SwitchedMultiHeadRRDB(SwitchedRRDB): + def __init__(self, nf, gc=32, num_convs=8, num_heads=2, init_temperature=1, final_temperature_step=1): + super(SwitchedMultiHeadRRDB, self).__init__(nf=nf, gc=gc, num_convs=num_convs, init_temperature=init_temperature, final_temperature_step=final_temperature_step) + self.RDB1 = SwitchedRDB_5C(nf, gc, num_convs=num_convs, num_heads=num_heads, init_temperature=init_temperature, collapse_heads=False) + self.RDB2 = SwitchedRDB_5C_MultiHead(nf, gc, num_convs=num_convs, num_heads=num_heads, init_temperature=init_temperature, collapse_heads=False) + self.RDB3 = SwitchedRDB_5C_MultiHead(nf, gc, num_convs=num_convs, num_heads=num_heads, init_temperature=init_temperature, collapse_heads=False) + self.RDB4 = SwitchedRDB_5C_MultiHead(nf, gc, num_convs=num_convs, num_heads=num_heads, init_temperature=init_temperature, collapse_heads=True) + + def set_temperature(self, temp): + [sw.set_attention_temperature(temp) for sw in self.RDB1.switches] + [sw.set_attention_temperature(temp) for sw in self.RDB2.switches] + [sw.set_attention_temperature(temp) for sw in self.RDB3.switches] + [sw.set_attention_temperature(temp) for sw in self.RDB4.switches] + + def forward(self, x): + out, att1 = self.RDB1(x, True) + out, att2 = self.RDB2(out, True) + out, att3 = self.RDB3(out, True) + out, att4 = self.RDB4(out, True) + + a1mean, _ = switched_conv.compute_attention_specificity(att1, 2) + a2mean, _ = switched_conv.compute_attention_specificity(att2, 2) + a3mean, _ = switched_conv.compute_attention_specificity(att3, 2) + a4mean, _ = switched_conv.compute_attention_specificity(att4, 2) + self.running_mean += (a1mean + a2mean + a3mean + a4mean) / 3.0 + self.running_count += 1 + + return out * 0.2 + x + # This module performs the majority of the processing done by RRDBNet. It just doesn't have the upsampling at the end. class RRDBTrunk(nn.Module): def __init__(self, nf_in, nf_out, nb, gc=32, initial_stride=1, rrdb_block_f=None, conv_first_block=None): @@ -295,21 +410,18 @@ class AssistedRRDBNet(nn.Module): return (out,) - class PixShuffleInitialConv(nn.Module): def __init__(self, reduction_factor, nf_out): super(PixShuffleInitialConv, self).__init__() self.conv = nn.Conv2d(3 * (reduction_factor ** 2), nf_out, 1) - self.r = reduction_factor + self.unshuffle = PixelUnshuffle(reduction_factor) def forward(self, x): (b, f, w, h) = x.shape # This module can only be applied to input images (with 3 channels) assert f == 3 - # Perform a "reverse-pixel-shuffle", reducing the image size and increasing filter count by self.r**2 - x = x.contiguous().view(b, 3, w // self.r, self.r, h // self.r, self.r) - x = x.permute(0, 1, 3, 5, 2, 4).contiguous().view(b, 3 * (self.r ** 2), w // self.r, h // self.r) - # Apply the conv to bring the filter account to the desired size. + + x = self.unshuffle(x) return self.conv(x) # This class uses a RRDBTrunk to perform processing on an image, then upsamples it. @@ -346,44 +458,4 @@ class PixShuffleRRDB(RRDBBase): fea = self.lrelu(self.upconv2(fea)) out = self.conv_last(self.lrelu(self.HRconv(fea))) - return (out,) - - -# This class uses two RRDB trunks to process an image at different resolution levels. -class MultiRRDBNet(RRDBBase): - def __init__(self, nf_base, gc_base, lo_blocks, hi_blocks, scale=2, rrdb_block_f=None): - super(MultiRRDBNet, self).__init__() - - # Chained trunks - lo_nf = nf_base * 4 - lo_nf_out = nf_base // 4 - hi_nf = nf_base - self.lo_trunk = RRDBTrunk(nf_base, lo_nf, lo_blocks, gc_base * 2, initial_stride=1, rrdb_block_f=rrdb_block_f, conv_first_block=PixShuffleInitialConv(4, lo_nf)) - self.skip_conv = nn.Conv2d(3, lo_nf_out, 1) - self.hi_trunk = RRDBTrunk(lo_nf_out, hi_nf, hi_blocks, gc_base, initial_stride=1, rrdb_block_f=rrdb_block_f) - self.trunks = [self.lo_trunk, self.hi_trunk] - - # Upsampling - self.scale = scale - self.upconv1 = nn.Conv2d(hi_nf, hi_nf, 5, 1, padding=2, bias=True) - self.upconv2 = nn.Conv2d(hi_nf, hi_nf, 5, 1, padding=2, bias=True) - self.HRconv = nn.Conv2d(hi_nf, hi_nf, 5, 1, padding=2, bias=True) - self.conv_last = nn.Conv2d(hi_nf, 3, 3, 1, 1, bias=True) - self.pixel_shuffle = nn.PixelShuffle(4) - - self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True) - - def forward(self, x): - fea_lo = self.lo_trunk(x) - fea = self.pixel_shuffle(fea_lo) + self.skip_conv(x) - fea = self.hi_trunk(fea) - - # Upsampling. - fea = F.interpolate(fea, scale_factor=2, mode='nearest') - fea = self.lrelu(self.upconv1(fea)) - if self.scale >= 4: - fea = F.interpolate(fea, scale_factor=2, mode='nearest') - fea = self.lrelu(self.upconv2(fea)) - out = self.conv_last(self.lrelu(self.HRconv(fea))) - return (out,) \ No newline at end of file diff --git a/codes/models/archs/arch_util.py b/codes/models/archs/arch_util.py index 9062bbf5..ecb7be76 100644 --- a/codes/models/archs/arch_util.py +++ b/codes/models/archs/arch_util.py @@ -128,3 +128,16 @@ def flow_warp(x, flow, interp_mode='bilinear', padding_mode='zeros'): vgrid_scaled = torch.stack((vgrid_x, vgrid_y), dim=3) output = F.grid_sample(x, vgrid_scaled, mode=interp_mode, padding_mode=padding_mode) return output + + +class PixelUnshuffle(nn.Module): + def __init__(self, reduction_factor): + super(PixelUnshuffle, self).__init__() + self.r = reduction_factor + + def forward(self, x): + (b, f, w, h) = x.shape + x = x.contiguous().view(b, f, w // self.r, self.r, h // self.r, self.r) + x = x.permute(0, 1, 3, 5, 2, 4).contiguous().view(b, f * (self.r ** 2), w // self.r, h // self.r) + return x + diff --git a/codes/models/networks.py b/codes/models/networks.py index 89bfc384..44ef9a94 100644 --- a/codes/models/networks.py +++ b/codes/models/networks.py @@ -38,14 +38,17 @@ def define_G(opt, net_key='network_G'): rrdb_block_f=functools.partial(RRDBNet_arch.SwitchedRRDB, nf=opt_net['nf'], gc=opt_net['gc'], init_temperature=opt_net['temperature'], final_temperature_step=opt_net['temperature_final_step'])) - elif which_model == 'MultiRRDBNet': - block_f = None - if opt_net['attention']: - block_f = functools.partial(RRDBNet_arch.SwitchedRRDB, nf=opt_net['nf'], gc=opt_net['gc'], - init_temperature=opt_net['temperature'], - final_temperature_step=opt_net['temperature_final_step']) - netG = RRDBNet_arch.MultiRRDBNet(nf_base=opt_net['nf'], gc_base=opt_net['gc'], lo_blocks=opt_net['lo_blocks'], - hi_blocks=opt_net['hi_blocks'], scale=scale, rrdb_block_f=block_f) + elif which_model == 'LowDimRRDBNet': + rrdb = functools.partial(RRDBNet_arch.LowDimRRDB, nf=opt_net['nf'], gc=opt_net['gc'], dimensional_adjustment=opt_net['dim']) + netG = RRDBNet_arch.RRDBNet(in_nc=opt_net['in_nc'], out_nc=opt_net['out_nc'], + nf=opt_net['nf'], nb=opt_net['nb'], scale=scale, rrdb_block_f=rrdb) + elif which_model == "LowDimRRDBWithMultiHeadSwitching": + switcher = functools.partial(RRDBNet_arch.SwitchedMultiHeadRRDB, num_convs=opt_net['num_convs'], num_heads=opt_net['num_heads'], + init_temperature=opt_net['temperature'], final_temperature_step=opt_net['temperature_final_step']) + rrdb = functools.partial(RRDBNet_arch.LowDimRRDBWrapper, nf=opt_net['nf'], gc=opt_net['gc'], dimensional_adjustment=opt_net['dim'], + partial_rrdb=switcher) + netG = RRDBNet_arch.RRDBNet(in_nc=opt_net['in_nc'], out_nc=opt_net['out_nc'], + nf=opt_net['nf'], nb=opt_net['nb'], scale=scale, rrdb_block_f=rrdb) elif which_model == 'PixRRDBNet': block_f = None if opt_net['attention']: